Terafab＋TSMC 合作構想分析報告
Terafab + TSMC Partnership Concept
Analysis Report
把巨大的算力需求轉成可管理的製造計畫
Turning Massive Compute Demand into a Manageable Manufacturing Plan
| 研究基準日：2026年10月6日 研究範圍：合作策略、技術分工、製造規劃、投資治理與執行風險。 |
Research reference date: October 6, 2026 Scope: partnership strategy, technology responsibilities, manufacturing planning, investment governance, and execution risks. |
| 閱讀說明：本報告以公開資料為基礎，將已披露資訊與建議方案分開呈現。合作架構、責任分工及投資門檻屬於分析建議，並非雙方已公布的合約內容。文中來源編號對應文末參考資料。 | Reading note: This report is based on public information and distinguishes disclosed facts from proposed arrangements. Partnership structures, responsibilities, and investment gates are analytical recommendations, not contractual terms announced by the parties. Source numbers refer to the references at the end. |
| 繁體中文 | English |
一、核心判斷
1. Core Assessment
| Terafab與台積電合作的最大價值，是把馬斯克企業的算力需求，轉換成可以設計、融資、製造、驗收與交付的晶片供應計畫。 | The greatest potential value of a Terafab–TSMC partnership is to translate the compute requirements of Musk-affiliated companies into a chip supply plan that can be designed, financed, manufactured, accepted, and delivered. |
| Terafab提供產品需求、系統應用與投資動力；台積電則可能提供製程、良率管理、量產營運及先進封裝能力。雙方若能建立明確的採購承諾與製造治理制度，可降低新建半導體產能的執行風險。 | Terafab brings product demand, system applications, and investment momentum. TSMC could provide process technology, yield management, volume-production operations, and advanced packaging capabilities. Clear purchase commitments and manufacturing governance could reduce the execution risks of new semiconductor capacity. |
| 截至研究基準日，馬斯克已確認雙方正在討論，但合作範圍、股權、製程及營運方式尚未定案。Culpium報導較可能的方向是台積電擁有並營運新廠，但其他安排仍在評估。[1] | As of the reference date, Musk had acknowledged discussions, but the scope, ownership, process technology, and operating model had not been finalized. Culpium reported that TSMC owning and operating the new factory was a likely direction, while alternatives remained under consideration. [1] |
| 主要判斷： 1. 先確定產品與訂單，再決定設備與廠房規模。 2. 製造責任必須與製程、品質及營運決策權相匹配。 3. 邏輯、記憶體、封裝與系統部署必須分別規劃，再共同整合。 4. 每期擴產應由驗證結果與實際需求推動。 |
Key assessments: 1. Define products and orders before sizing equipment and facilities. 2. Manufacturing accountability must match decision rights over process, quality, and operations. 3. Plan logic, memory, packaging, and system deployment separately, then integrate them. 4. Base each expansion phase on validation results and actual demand. |
二、公開資訊與證據邊界
2. Public Information and Evidence Limits
| 事項｜公開資訊｜尚待確認 | Item | Public information | Unresolved questions |
| TSMC合作討論｜馬斯克確認正在討論｜正式協議、投資額、責任範圍。[1] | TSMC discussions | Musk acknowledged talks | Formal agreement, investment amount, and responsibilities remain unresolved. [1] |
| 台積電德州布局｜Reuters報導正在評估，尚未定案｜董事會核准、選址、產能及時程。[2] | TSMC Texas expansion | Reuters reported an evaluation, with no finalized plan | Board approval, location, capacity, and timing remain unresolved. [2] |
| Terafab第一期｜德州政府公布Grimes County第一期投資超過168億美元，預計創造3,000個職位｜實際支出、設備進度及量產結果。[3] | Terafab Phase 1 | Texas announced more than US$16.8 billion of investment in Grimes County and 3,000 planned jobs | Actual spending, equipment progress, and production results require confirmation. [3] |
| 多階段園區計畫｜JETI申請文件提出最多四階段，投資估計550億至1,190億美元｜各階段是否執行、最終預算。[4] | Multiphase campus | A JETI application proposed up to four phases with estimated investment of US$55–119 billion | Execution of each phase and final budgets remain uncertain. [4] |
| Intel參與｜Intel官方披露已加入Terafab成為策略夥伴｜與台積電的分工及最終製程安排。[5] | Intel participation | Intel officially disclosed joining Terafab as a strategic partner | Responsibilities relative to TSMC and final process arrangements remain unresolved. [5] |
| 投資估計、地方協議與合作討論，都不能直接視為已完成的製造能力。JETI申請內容為申請方提出的規劃，並非獨立驗證的完工或量產結果。 | Investment estimates, local agreements, and partnership discussions do not establish completed manufacturing capacity. The JETI application describes the applicant’s plans, not independently verified construction or production outcomes. |
| DIGITIMES標題中的「超過80%」反映供應商對合作的看法。可讀內容未提供估算方法與受訪樣本，本報告不將其用作財務模型中的成交機率。[6] | The “over 80%” figure in the DIGITIMES headline reflects suppliers’ views. The accessible content did not disclose the estimation method or respondent sample; this report does not use it as a transaction probability in financial modeling. [6] |
三、Terafab的戰略定位
3. Terafab’s Strategic Positioning
| Terafab官網提出整合邏輯、記憶體及先進封裝，並服務車用、機器人與太空運算等用途。德州申請文件進一步描述設計、光罩、晶圓製造、封裝與系統整合的共同園區構想。[4][7] | Terafab’s website envisions integrating logic, memory, and advanced packaging for applications including vehicles, robotics, and space computing. The Texas application further describes a campus combining design, photomasks, wafer fabrication, packaging, and system integration. [4][7] |
| 層次｜核心問題｜成功條件 | Layer | Key question | Success requirement |
| 產品｜要製造哪些晶片？｜規格明確、設計可驗證。 | Product | Which chips should be manufactured? | Clear specifications and verifiable designs. |
| 製造｜如何穩定交付合格晶片？｜製程、良率、封裝與測試到位。 | Manufacturing | How can qualified chips be delivered consistently? | Ready process technology, yield, packaging, and testing. |
| 商業｜新增產能由誰使用與支付？｜訂單、資金與成本分攤確定。 | Commercial | Who will use and pay for new capacity? | Committed orders, funding, and cost allocation. |
| 分析判斷：台積電若加入，能強化製造層的執行能力；產品是否有足夠需求，以及系統能否按時部署，仍需要Terafab及其客戶負責落實。 | Analytical assessment: TSMC’s participation could strengthen manufacturing execution. Terafab and its customers would still need to establish sufficient product demand and deliver systems on schedule. |
四、把算力需求轉換成製造需求
4. Translating Compute Demand into Manufacturing Requirements
4.1 分別建立產品需求 / Define Demand by Product
| 各類應用不宜共用單一需求預測。 | Different applications should not share a single undifferentiated demand forecast. |
| 產品方向｜需求規劃｜製造與驗證 | Product category | Demand planning | Manufacturing and validation |
| 車用運算｜車輛數量、每車晶片配置、產品週期｜品質追溯、可靠度與供貨延續。 | Vehicle computing | Vehicle volumes, chips per vehicle, product lifecycle | Traceability, reliability, and supply continuity. |
| 機器人運算｜部署數量、感知與控制需求、單機成本｜功耗、尺寸、散熱與迭代速度。 | Robot computing | Deployment volumes, perception and control requirements, unit cost | Power, size, thermal design, and iteration speed. |
| 地面AI運算｜訓練／推論工作負載、系統數量｜記憶體頻寬、封裝與系統整合。 | Terrestrial AI computing | Training/inference workloads and system volumes | Memory bandwidth, packaging, and system integration. |
| 太空運算｜任務數量、能源預算、部署條件｜任務環境、容錯與可靠度驗證。 | Space computing | Mission volumes, energy budgets, deployment conditions | Mission environment, fault tolerance, and reliability validation. |
| 以上為規劃框架，尚非Terafab已確認的產品BOM。 | This is a planning framework, not a confirmed Terafab product bill of materials. |
4.2 以合格晶片反推投片量 / Work Back from Qualified Chip Demand
| 年度成品晶片需求＝部署設備數量×每台晶片數量＋備品與維修需求。 | Annual finished-chip demand = deployed equipment units × chips per unit + spare and replacement demand. |
| 單一晶粒產品的年度晶圓需求＝年度合格成品需求÷（每片晶圓理論晶粒數×晶粒良率×後段綜合良率）。 | For a single-die product: annual wafer demand = annual qualified finished-chip demand ÷ (theoretical dies per wafer × die yield × combined downstream yield). |
| 多晶粒產品需分別計算每種晶粒，再檢查封裝、記憶體與測試能力是否足以配套。後段綜合良率應涵蓋封裝與最終測試。 | For multidie products, calculate each die type separately and check whether packaging, memory, and testing capacity can support the resulting volumes. Combined downstream yield should cover packaging and final testing. |
| 製造計畫的最終單位，應是可驗收的合格出貨量。單靠FLOPS、用電容量或廠房面積，無法可靠決定投片與設備需求。 | The ultimate planning measure should be qualified shipments that meet acceptance criteria. FLOPS, electrical capacity, or building area alone cannot reliably determine wafer starts and equipment requirements. |
五、建議合作架構
5. Proposed Partnership Structure
5.1 台積電主導製造，客戶參與出資 / TSMC-Led Manufacturing with Customer Investment
| 台積電已有製造商控股、客戶參股的JASM先例。2024年擴建公告列出的交易完成後股權安排為台積電約86.5%，其餘由Sony、DENSO與Toyota持有。此先例可供治理設計參考，股權比例與製程條件不能直接套用。[8] | JASM provides a precedent for a manufacturer-controlled venture with customer investment. Its 2024 expansion announcement specified approximately 86.5% ownership for TSMC after completion of the investment, with the remainder held by Sony, DENSO, and Toyota. The governance concept is relevant, but its ownership ratios and process conditions cannot be directly applied. [8] |
| 責任方｜建議承擔事項 | Party | Proposed responsibilities |
| Terafab及需求企業｜產品規格、設計交付、資金、預付款與最低採購承諾。 | Terafab and demand-side companies | Product specifications, design delivery, funding, prepayments, and minimum purchase commitments. |
| 台積電｜製程導入、設備驗證、製造管理、品質及量產放行。 | TSMC | Process deployment, equipment qualification, manufacturing management, quality, and production release. |
| 記憶體與後段供應商｜合格料件、封裝／測試產能、可靠度與交期。 | Memory and downstream suppliers | Qualified components, packaging/testing capacity, reliability, and delivery schedules. |
| 共同治理團隊｜預算、需求變更、風險分攤及擴產核准。 | Joint governance team | Budgets, demand changes, risk allocation, and expansion approvals. |
5.2 合約必須解決的五項問題 / Five Contractual Issues
| 需求縮減：客戶取消或延後採購時，如何補償已投入的產能？ 設計延遲：設備已就緒、產品未完成時，閒置成本由誰承擔？ 良率不足：製程、設計及封裝問題如何判定責任？ 成本增加：材料、人工與公用設施成本如何反映於價格？ 交付優先權：保留產能與其他客戶訂單如何協調？ |
Demand reduction: How is committed capacity compensated when purchases are canceled or deferred? Design delays: Who bears idle-capacity costs when equipment is ready but designs are not? Yield shortfalls: How is responsibility assigned among process, design, and packaging issues? Cost increases: How are materials, labor, and utility costs reflected in prices? Delivery priority: How is reserved capacity coordinated with other customer orders? |
| 合作的經濟價值，將高度取決於這些條款，而非單看投資總額。 | The partnership’s economic value will depend heavily on these terms, rather than on the headline investment amount alone. |
六、技術與供應鏈分工
6. Technology and Supply Chain Responsibilities
| 台積電官方技術路線涵蓋先進邏輯、CoWoS、SoIC與COUPE光學整合，可提供多種技術選項；目前尚不能據此認定Terafab會採用哪一項。[9] | TSMC’s official roadmap covers advanced logic, CoWoS, SoIC, and COUPE optical integration, offering multiple technology options. This does not establish which technologies Terafab will adopt. [9] |
| 環節｜應確認內容｜主要風險 | Area | Required clarification | Main risk |
| 邏輯製程｜節點、設計平台、導入地點與時程｜設計與製程成熟度不匹配。 | Logic process | Node, design platform, deployment location, and timing | Mismatch between design readiness and process maturity. |
| 記憶體｜類型、數量、供應商與採購承諾｜運算晶粒完成但記憶體不足。 | Memory | Type, volumes, suppliers, and purchase commitments | Compute dies ready but insufficient memory supply. |
| 先進封裝｜平台、材料、產能與驗證責任｜後段能力限制成品出貨。 | Advanced packaging | Platform, materials, capacity, and validation responsibilities | Downstream capacity constrains finished shipments. |
| 光學互連｜系統需求與導入效益｜過早增加設計及驗證複雜度。 | Optical interconnects | System requirements and adoption benefits | Premature increases in design and validation complexity. |
| 測試｜晶圓、封裝與系統測試分工｜缺陷未在適當階段被發現。 | Testing | Wafer, package, and system-test responsibilities | Defects escape detection at the appropriate stage. |
| 廠務｜電力、水、氣體及維護能力｜設備到位但公用設施不完整。 | Facilities | Power, water, gases, and maintenance capabilities | Equipment arrives before supporting utilities are ready. |
| 建議先完成可量產的基礎方案，再依產品需求引入更複雜的整合技術。 | First establish a manufacturable baseline, then introduce more complex integration technologies as justified by product requirements. |
七、Intel的角色
7. Intel’s Role
| Intel官方已確認參與Terafab。台積電的合作討論，尚不足以證明Intel已被取代。[5] | Intel has officially confirmed its participation in Terafab. Discussions with TSMC do not establish that Intel has been replaced. [5] |
| 可研究的安排包括：不同產品由不同供應商製造、分階段導入不同產線，或不同晶粒分別製造後整合。 | Potential arrangements include manufacturing different products with different suppliers, introducing production lines in stages, or manufacturing different dies separately before integration. |
| 若採多供應商模式，必須按產品、晶粒與製造環節明確分配責任。供應來源增加的同時，也會增加設計、測試及交付協調工作。 | A multisupplier model requires explicit responsibilities by product, die, and manufacturing stage. More supply sources also increase design, testing, and delivery coordination requirements. |
八、投資與財務管理
8. Investment and Financial Management
| 目前資料不足以可靠計算新增營收、IRR、NPV或回收年限。後續應分別建立台積電製造模型與Terafab成品模型。 | Available data are insufficient to reliably calculate incremental revenue, IRR, NPV, or payback. Separate TSMC manufacturing and Terafab finished-product models should be developed. |
| 模型｜主要輸入 | Model | Key inputs |
| 台積電製造模型｜投片量、利用率、計價方式、成本、折舊及爬坡支出。 | TSMC manufacturing model | Wafer starts, utilization, pricing terms, costs, depreciation, and ramp-up spending. |
| Terafab成品模型｜晶粒數、良率、記憶體、封裝、測試與合格成品數量。 | Terafab finished-product model | Die counts, yields, memory, packaging, testing, and qualified finished volumes. |
| 投資模型｜出資比例、預付款、融資、補助、營運資金及擴產支出。 | Investment model | Equity contributions, prepayments, financing, incentives, working capital, and expansion spending. |
| 晶粒良率直接影響取得合格晶片的成本；其對代工營收的影響，需依計價與責任條款判斷。 | Die yield directly affects the cost of obtaining qualified chips. Its impact on foundry revenue depends on pricing and contractual accountability. |
| 情境｜主要假設｜評估目的 | Scenario | Main assumptions | Purpose |
| 保守｜需求較低、量產延遲、爬坡支出較高｜測試資金承受能力。 | Conservative | Lower demand, production delays, higher ramp-up spending | Test financial resilience. |
| 基準｜依承諾訂單與驗證後成本建模｜評估可執行性。 | Base case | Committed orders and validated costs | Assess executable economics. |
| 擴張｜訂單增加、利用率提高、效率改善｜評估後續擴產價值。 | Expansion | Higher orders, utilization, and efficiency | Evaluate subsequent expansion value. |
九、分階段執行計畫
9. Phased Execution Plan
| 階段｜主要工作｜進入下一階段條件 | Stage | Main work | Gate for progression |
| 需求確認｜建立產品、數量及部署計畫｜規格與採購需求獲確認。 | Demand definition | Define products, volumes, and deployment plans | Specifications and purchasing requirements confirmed. |
| 設計與試產｜完成設計、製程及封裝驗證｜性能、品質與成本符合門檻。 | Design and pilot production | Validate design, process, and packaging | Performance, quality, and costs meet agreed thresholds. |
| 量產準備｜設備、廠務、人員與供應鏈到位｜完成量產資格驗證。 | Production readiness | Prepare equipment, facilities, personnel, and supply chain | Volume-production qualification completed. |
| 初期量產｜穩定良率、產出及交付｜合格出貨支持客戶部署。 | Initial volume production | Stabilize yield, output, and delivery | Qualified shipments support customer deployment. |
| 後續擴產｜增加設備與產能｜訂單、利用率及現金流支持投資。 | Subsequent expansion | Add equipment and capacity | Orders, utilization, and cash flow justify investment. |
| 本報告不預設統一良率門檻或量產日期。這些條件應由產品規格、試產數據及合約共同決定。 | This report does not prescribe universal yield thresholds or production dates. These should be determined jointly by product specifications, pilot data, and contracts. |
十、管理指標與主要風險
10. Management Metrics and Key Risks
| 指標｜管理用途 | Metric | Management purpose |
| 訂單覆蓋率｜檢查新增產能受到多少採購承諾保障。 | Order coverage | Assess how much new capacity is backed by purchase commitments. |
| 設計里程碑達成率｜判斷產品能否按時試產。 | Design milestone attainment | Assess readiness for on-time pilot production. |
| 合格出貨量｜衡量可實際使用的晶片供應。 | Qualified shipment volume | Measure usable chip supply. |
| 每顆合格晶片成本｜追蹤良率、材料及後段成本。 | Cost per qualified chip | Track yield, materials, and downstream costs. |
| 準時交付率｜檢查是否支持部署時程。 | On-time delivery | Assess support for deployment schedules. |
| 資本支出與現金流｜控制投資速度與資金需求。 | Capital spending and cash flow | Control investment pace and funding needs. |
| 最主要的風險，是需求願景成熟速度慢於設備投資速度。其次為製造權限不清、記憶體與封裝能力不足，以及多產品、多供應商帶來的整合負擔。 | The principal risk is that demand matures more slowly than equipment investment proceeds. Additional risks include unclear manufacturing authority, insufficient memory and packaging capacity, and integration burdens across multiple products and suppliers. |
| Terafab＋TSMC值得持續研究的合作方向，是以可靠訂單支撐產能，以製造治理保障交付，以階段驗證控制投資。正式評估合作價值前，仍需取得製程、產能、採購承諾、出資及營運權等關鍵條款。 | The partnership merits continued study as a model in which reliable orders support capacity, manufacturing governance protects delivery, and staged validation controls investment. A definitive assessment requires terms covering process technology, capacity, purchase commitments, funding, and operating authority. |
參考資料
References
[1] Culpium — Elon Musk Confirms TSMC-Terafab Talks
2026-10-03
https://www.culpium.com/p/elon-musk-confirms-tsmc-terafab-talks
[2] Reuters — TSMC evaluates potential Texas investment, sources say
2026-09-30; updated 2026-10-01
https://www.reuters.com/world/asia-pacific/tsmc-evaluates-potential-texas-investment-sources-say-2026-09-30/
[3] Office of the Texas Governor — Governor Abbott Announces SpaceX Expansion In Grimes County
2026-08-06
https://gov.texas.gov/news/post/governor-abbott-announces-spacex-expansion-in-grimes-county
[4] Texas Comptroller — J0036 Terafab Anderson-Shiro JETI application; project description, PDF p. 7
2026 application / 2026年申請文件
https://assets.comptroller.texas.gov/open-data/jeti/J0036/J0036-terafab-anderson-app.pdf
[5] Intel — First-Quarter 2026 Financial Results, SEC filing
2026-04-23
https://www.sec.gov/Archives/edgar/data/50863/000005086326000077/q126earningsrelease.htm
[6] DIGITIMES — TSMC emerges as potential Terafab partner as suppliers see over 80% chance of Musk deal
2026-10-05; full text subscription-only / 全文需訂閱
https://www.digitimes.com/news/a20261005VL204/tsmc-fab-manufacturing-wafer-intel.html
[7] Terafab — Official project website / 官方專案網站
Reviewed 2026-10-06 / 查核2026-10-06
https://www.terafab.ai/
[8] TSMC — JASM Set to Expand in Kumamoto Japan
2024-02-06
https://pr.tsmc.com/english/news/3105
[9] TSMC — TSMC Debuts A13 Technology at 2026 North America Technology Symposium
2026-04-23
https://pr.tsmc.com/english/news/3302